Image Multiscale Decomposition Based On Neighbor Distance And Its Application

Multi-scale structure is one of the elemental qualities of natural image. Thelarge-scale sub-image reflects the source image’s structure， contains lots of lowfrequency information， and can be regarded as a low frequency component of thesource image; The small-scale sub-image contains lots of detail and textureinformation， and can be regarded as a high frequency component of the source image.Many image multi-scale decomposition methods that can effectively extract the image’smulti-scale information have been widely and successfully used in various imageprocessing tasks， such as， image denoising， image segment， image claritymeasurement and multi-sensor image fusion.In the multi-scale Laplacian decomposition scheme， the high frequencycomponent is the difference of the source image and its low frequency component.Therefore， the high frequency is not accurate when they are used to measure thevariance of pixel gray values. In the view of the high frequency component’simportance in many image processing tasks， we construct a new image high frequencyfilter， propose a new image multi-scale decomposition method， and then apply theproposed decomposition method to no-reference image sharpness measurement andmulti-sensor image fusion in this dissertation.The main contributions of this dissertation are summarized as follows:（1） In the view of the effectiveness of oriented distance which measure thedistance of two points on a smooth surface，we firstly restore the smooth image surfaceby non-parametric regression， secondly construct a neighbor distance filter to extractthe image’s high frequency component， and finally proposed a new image multi-scaledecomposition scheme based on the neighbor distance filter.（2） A fast and effective sharpness estimation of an image is necessary forauto-focus in optical imaging system and automatic drawing the image region ofinterest. As the image’s sharpness information exist in the high frequency component，we use the neighbor distance filter proposed in chapter2to extract the image’s highfrequence component， and use the extracted high frequency component to construct anew and fast no-reference image sharpness index. The result of experiments shows thehigh performance of the proposed sharpness index.（3） The effective selection of sharp image blocks is a key factor in multi-focus image fusion. In chapter3， we firstly divide every source images into blocks，secondly compute the sharpness of each image blocks， and finally compare thesharpness of each corresponding blocks and construct the block of the composite imageby choosing the sharpest image blocks. The experiments demonstrate that the proposedfusion method is superior to the conventional image fusion methods in terms of someobjective evaluation indexes.（4） To overcome the difficulty of selecting the best block’s size and the blockeffects in the result image， we firstly decompose every source image into a series ofhigh frequency components and a low frequency component by the multi-scaledecomposition method proposed in chapter2， secondly compute the pixel contrast bythe ratio of the high frequency component coefficients to the low frequency componentcoefficients， and finally choose the high and low frequency component coefficients ofsource pixel with biggest pixel contrast and reconstruct the fused image by imagereconstruction algorithm. The experiments show that the image fusion method based onthe pixel contrast can generate a better fusion image than the conventional image fusionbased on multi-scale image decomposition methods in terms of some objectiveevaluation indexes.（5） Besides singular points， the natural image also contains many singular linewith different direction. To overcome the drawback that the multi-scale imagedecomposition method based on the neighbor distance only extract the singular pointinformation， we combine the directional filter banks with the proposed multi-scaledecomposition method， construct a multi-directional and multi-scale imagedecomposition method， and use the proposed image decomposition method to fusemulti-sensor images. Experiment results on120multi-sensor images show that theproposed fusion method achieves better results that conventional image fusion methodsin terms of some objective evaluation indeses.